Which AI Makes the Best Marketing Websites?

Welcome back to The Signal, a weekly letter where I share stories, trends, strategies and insights to help you level up as a creator and entrepreneur.

Today is the first episode of my new series, Agent Wars – a head-to-head competition where leading AI agents tackle the same challenge to see which model is most capable.

Make sure to subscribe to get future episodes delivered right to your inbox.

The Challenge

Five leading AI models receive the same three website briefs and are blindly judged on functionality, taste, and creativity.

How It Works

For Episode 01, here are the models I tested, with their reasoning level and the harness I used for each:

• GPT 5.6-Sol (High) on the ChatGPT desktop app

• Fable 5 (High) on the Claude desktop app

• Grok 4.5 (High) on Cursor

• Gemini 3.5 Flash (High) on Antigravity desktop app

• Muse Spark 1.1 (Thinking) on meta.ai

Each agent received the same three creative briefs:

• Build a world-tour website for a fictional band

  • Invent a fictional band, tour identity, global schedule, ticket system, discography, merchandise, and complete visual world.

• Create the launch site for an original AAA fantasy RPG

  • Create an original game universe with characters, factions, gameplay pillars, editions, world map, countdown, and speculative imagery.

• Design a premium coffee shop and co-working website

  • Follow a tightly constrained coffee-and-coworking brief where names, colors, prices, hours, menu items, and membership details had to be exact.

Every build was judged across three categories: technical execution, taste, and creativity. That included 25 objective technical checks per website—covering responsiveness, accessibility, performance, reliability, functionality, and content accuracy — plus separate scores for visual polish and originality.

Each round was worth 50 points, for a possible total of 150.

Here is a link to the prompts, scoring rubicons, estimated cost and time of each build, and their final scores.

Scoring

Each website was scored in three separate dimensions:

The Rubicon: 25 Points

Objective pass-or-fail checks covering delivery, content accuracy, controls, responsive behavior, stability, and performance.

Creativity: 15 Points

Central concept, originality, worldbuilding, interaction design, depth, and commitment.

Taste: 10 Points

Composition, typography, color, visual craft, cohesion, restraint, and polish.

Across three rounds, each model could earn 150 points. Time and total cost were tracked separately for reference and did not affect the competition score.

Results

GPT 5.6-Sol: 121

Fable 5: 118

Muse Spark 1.1: 100

Grok 4.5: 99

Gemini 3.5 Flash: 99

OpenAI’s GPT 5.6-Sol won Episode 01, but Claude finished just three points behind, while the remaining three competitors were separated by only a single point.

Notes

A few minor distinctions between the models to take note of:

• Unlike the other models, Anthropic has no native media generation capabilities. I connected the Higgsfield MCP (which I covered here) for media generation.

• Meta’s Muse Spark 1.1 is the only model without its own harness. I interfaced with it directly on meta.ai. Choosing the thinking model default to Muse Spark 1.1.

Final Thoughts

I used the winning model of this episode, GPT 5.6-Sol, to create the Agent Wars website.

It took under 30 minutes to create and deploy.

Currently, as of July 2026, the very best models are currently still ways away from the human craft and precision of the world’s best web builders.

But my estimate is that the frontier models are around ~65%-70% of the way there.

Which is incredible. Because you get a pretty damn good website for a few dollars of compute in just a few minutes.

That’s completely world-changing for most individuals and small businesses. There really is no excuse anymore to have a subpar website. You can literally paste your website into Claude or ChatGPT and enter a simple prompt like the following:

Roberto Nickson

Say hi on X, Instagram or YouTube